How do you think the new GigE standards will influence the machine vision industry?
Respond or ask your question now!
By Keith Reid
The biosciences and medicine also provide some significant challenges for automated image analysis. Although the environment, is fairly controlled and the elements being analyzed should have some degree of commonality, there are numerous different types of cells, numerous types of diseases to be studied, natural diversity among cells of the same type and often lighting and image quality challenges.
Lifescan Biosciences (Seattle, Wash.) focuses on developing image analysis software to analyze pathology specimens. This involves four-micron sections of human or animal tissue that have been fixed and embedded in wax then placed on glass slides. As has been common for the past 100 years, the specimens are then typically stained with two dies—a one pink and one blue. Analysis can take the form of cytology, which involves looking at single cells and analyzing the size, shape and other properties of that object. It can also involve histology, which is looking at a section of tissue containing multiple cells and histopathology where the focus is on disease.
"With histology the first challenge is to determine which objects are being analyzed, and one useful place to start is with the nuclei," said Joseph Brown, PhD., Lifescan president and CEO. "Nuclei are oval objects in a fairly well-defined size range in microns. Most cells have a single nucleus so it's a good place to start and they stain blue. You have some immediate problems. Because you are looking at a section that is four microns thick you are often looking at nuclei that have been sliced in half, which creates a sampling problem which makes measuring the size of the nuclei more problematical. Once you've found nuclei you can analyze the properties of the nuclei of the size shape and texture using various mathematical tools to address particular tasks."
Lifecan developed image analysis software for a major program aimed at detecting cancer cells from normal cells, by examining the nuclei to identify internal features that statistically distinguish cancer from normal cells. By combining a sufficient number of these features accuracies approaching 99 Percent can be achieved using this "IDG" approach. Another approach being explored by Lifescan in conjunction with NEC of Japan is "Web Fractal," which involves examining the relationships between the separate nuclei. "Normal tissues tend to be very well structured," said Brown. "Cancer, by its very nature, is a disorganization of that highly ordered normal structure and a tumor means a lump. So cancers tend to lumps of tissue that are disorganized."
The Web Fractal approach is also being applied to toxicologic pathology, which involves testing drugs on animals and then analyzing the tissue samples from the test and control groups to note any tissue changes. "At first glance, replacing a human pathologist is a major challenge and using a machine to make a clinical diagnosis to say that this specimen is a certain type of breast cancer with a tumor grade and a stage is not our immediate target," said Brown. "The target is more realistically to look in the specimen and get a "yes/no" answer as to whether or not it is normal, and what is the probability of it being cancer. You can either use this technology to review specimens before the human looks at them to draw his or her attention to certain features, or you can use it to look at specimens after the human is done to make sure that nothing was missed. Both are valid approaches."